Jeff Dean, the "founding infrastructure veteran" at Google, has left the company to launch his own startup, which targets AI4S.

In 1999, when Jeff Dean joined Google, the company had only 20 employees, with its office located on the upper floor of a T-Mobile store in Palo Alto. 27 years later, he is Google's Chief Scientist, and behind him are MapReduce, Bigtable, Spanner, TensorFlow, TPU, and Gemini — almost every core infrastructure that supports Google's operations bears his mark.

On August 5, Jeff Dean announced his departure from Google to co-found Discovery Loop with three long-term collaborators. On the day the news was released, Alphabet's share price fell by about 5%.

The market's reaction is a clear reflection of reality: it is not just an executive leaving, but a load-bearing pillar of the company.

01 Building the Infrastructure for AI Research

The goal of Discovery Loop, in the words of its founding team, is to "automate the experimental loops of machine learning, science and engineering".

This statement needs to be broken down to understand. The core process of scientific research is essentially a cycle — putting forward hypotheses, designing experiments, executing experiments, verifying results, then proposing new hypotheses based on the results and starting a new round. Humans have been conducting scientific research for hundreds of years relying on this cycle. The problem is that this cycle is extremely dependent on manual work, and its speed is limited by human energy and time.

What Discovery Loop aims to do is use AI models and large-scale computing power to automate this cycle and run thousands of experiments in parallel.

All members of the Discovery Loop team are top technical talents from Google | Image source: x

The specific roadmap is divided into three steps:

First step, start with machine learning research itself — use AI to conduct AI research automatically, automatically propose experimental schemes, run evaluations, and iterate models. This is the most pragmatic starting point, because the feedback loop of ML research is the shortest, the evaluation indicators are the clearest, and the data is the most readily available.

Second step, use this set of capabilities to optimize its own technology stack, making itself its first customer.

Third step, expand to broader fields of science and engineering, including hardware design, drug discovery and clean energy.

The company is registered as a Public Benefit Corporation, following the same governance path as OpenAI, Anthropic, xAI, and Ilya Sutskever's SSI.

The company's seed round is jointly led by Radical Ventures and Khosla Ventures, with participation from Lightspeed, Kleiner Perkins, Doerr Capital, and Alphabet itself. The amount and valuation have not been disclosed.

It is worth noting that Google not only invested capital, but also will serve as a cloud partner to provide Discovery Loop with computing power support for at least the first year.

This detail is thought-provoking — the old employer is willing to allocate computing power resources to support an external company founded by its core employees, which is not common in Silicon Valley.

02 "Living Fossil of Infrastructure"

The founding team of Discovery Loop has only four members, but the sum of these four people is almost a brief history of modern computing and AI infrastructure.

Jeff Dean, Google's 30th employee, a 27-year veteran. Together with Sanjay Ghemawat, he created MapReduce, Bigtable, Google File System and Spanner — these are not fancy application-layer products, but the underlying "operating system" that enables Google to process data on a global scale. Without these, there would be no Google Search, no Gmail, no Google Ads. Later, he led Google Brain, promoted the development of TensorFlow and the design of TPU chips, and served as Chief Scientist after the merger of Google Brain and DeepMind in 2023.

Sanjay Ghemawat, Google Senior Fellow, Dean's partner for more than 20 years. Inside Google, Dean and Ghemawat are a legendary duo — the two almost always appear together in the most important papers and projects. Some people joke that they share one brain.

Oriol Vinyals, Vice President of Research at Google DeepMind, one of the technical leads of Gemini. During his time at DeepMind, he participated in landmark projects such as AlphaStar (the AI that defeated professional StarCraft players) and AlphaCode (code generation).

Quoc Le, co-founder of Google Brain, one of the pioneers of AutoML, who made foundational work in the field of sequence-to-sequence learning.

Jeff Dean's departure announcement posted on x | Image source: x

Content of Jeff's departure announcement:

Tomorrow will be my last day at Google. After 27 years here, I have witnessed it grow from 25 people to more than 190,000 people. It has been an amazing journey. Below is a note I shared with many people inside Google today. Excerpts are as follows:

It has been an absolute privilege to work with you and help build some of the most widely used and impactful products ever made. When I was a kid, I dreamed of helping build software that would be used by many people, and now Google has 13 products used by more than a billion people (incredible!).

Our work has had a huge impact on the world, and I have been fortunate to work with and build friendships with many colleagues whom I deeply admire, respect, and enjoy being around. Every time I see people all over the world using our products to find information, process emails, translate documents, watch videos, learn new things, navigate and understand the physical world, browse the web, use mobile phones, run large-scale computing on our infrastructure, ride autonomous vehicles, or perform complex tasks with the help of our AI systems, I still feel extremely joyful. I hope you all can share this joy, because it is a collective achievement! Thank you to all my colleagues at Google over the years!

Now I am very excited to start @DiscoLoopAI with my old friends and colleagues @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix.

The four people have worked together for periods ranging from 14 to 30 years. The introduction on Discovery Loop's official website states that this team includes three of the most cited researchers in the AI field and two of the most cited researchers in the distributed systems field.

If you look closely at this resume, you will find a keyword that runs through Dean's entire career — infrastructure. What he did at Google was never user-facing products, but the layer that systematically amplifies other people's work. MapReduce amplified engineers' ability to process data, TensorFlow amplified researchers' ability to train models, and TPU amplified the computing power supply of the entire industry.

Now, what he wants to amplify is the ability of scientists to conduct research. The tools have changed, but the role remains the same. The essence of Discovery Loop is "infrastructure for scientific discovery".

03 The AI Migration Wave

Jeff Dean's departure is not an isolated event.

This summer alone, Google DeepMind has experienced an unprecedented brain drain. In June, Nobel laureate John Jumper (core of the AlphaFold team) left to join Anthropic, and Noam Shazeer, co-author of the Transformer paper, also left not long ago. In just one week in June, five senior DeepMind researchers left for Anthropic and OpenAI respectively, and Alphabet's market value once evaporated by about 270 billion US dollars.

Now, even the people who built MapReduce have left.

And this is not just Google's problem. Mira Murati left OpenAI in 2024 and founded Thinking Machines Lab, raising 2 billion US dollars in financing within 5 months with a valuation of 12 billion US dollars. After leaving OpenAI, Ilya Sutskever founded SSI, focusing on superintelligence safety. Andrej Karpathy also joined Anthropic this year.

A thought-provoking trend is emerging — the most core technical talents in the AI industry are systematically migrating from large companies to startups.

The reason behind this is not just money. The pre-IPO equity of Anthropic and OpenAI does have huge appeal, but Dean does not lack money at Google. The reason he gave himself is more noteworthy — The quarterly earnings pressure of public companies limits the freedom of scientific decision-making. The subtext of this sentence is that in a system that needs to account to Wall Street for quarterly revenue, it is difficult to carry out basic research on a ten-year scale.

The fact that Discovery Loop is registered as a public benefit corporation essentially solves this problem at the institutional level. The PBC structure allows the company to maintain a legally protected mission commitment while pursuing commercial returns, and will not be forced to give up long-term scientific goals for short-term monetization due to shareholder pressure.

Google is obviously aware of this contradiction. Alphabet itself invested in Discovery Loop and provided computing power for the first year — this is almost an admission that some things cannot be done within the company, but it does not want to lose these people completely. This is a pragmatic choice for a large company facing talent outflow, but it also exposes a structural dilemma.

At the same time, Google is also making internal adjustments. In the restructuring announced on the same day as Dean's departure, Demis Hassabis stepped down from the day-to-day management of Google DeepMind and became Chairman and Chief Scientist of Alphabet; CTO Koray Kavukcuoglu was promoted to Senior Vice President to take over the development of the Gemini model. Pichai said in an internal letter that this is to inject new momentum into Google's AI work.

But the signal the market reads may be just the opposite.

Analysts noted that the number of iconic figures leaving Google's AI department this summer has exceeded any comparable period in the company's history. It is not "the people who train models on the infrastructure" that left, but "the people who build the infrastructure itself". This is a qualitative change.

Jeff Dean said a sentence at the University of Washington's graduation ceremony in June this year, which now looks like a preview in hindsight. He told computer science graduates that one of the problems worth solving is "developing tools that accelerate scientific discovery and engineering".

He also said another sentence that probably better explains why he left.

"The beauty of software is that a small group of people can build something that has a huge impact on the world."

27 years ago, he proved this in a 20-person startup. Now, at the age of 58, he wants to prove it again.

This article is from the WeChat public account "GeekPark" (ID: geekpark), author: Hualin Wuwang, editor: Jingyu, published with authorization from 36Kr.